An artificial neural network model for river flow forecasting

Abstract

An artificial neural network model developed to forecast river flow/discharge is presented. The model is based on the feed-forward, back-propagation network architecture optimized through a conjugated training algorithm. The observations of daily mean river discharge and daily mean rainfall data during the period 2002 to 2005 were used to train the model and verify the model predictions. The Hanwella station downstream of the Kelani river basin was chosen as the target station where the discharge was to be predicted. The river discharge at the Glencourse, Deraniyagala, Holombuwa and Kithulgala stations upstream of the target station were selected as the main inputs to the model. The contribution of inputs to the model output was determined by studying the correlation between the predicted and the actual discharge values. The best correlation coefficients between the forecasted and observed discharge for the Hanwella station are 0.95, 0.86 and 0.68 for the same day, one day ahead and two days ahead forecasting respectively